Brain Models - What will be next?

Clark's article http://dericbownds.net/uploaded_images/Clark_preprint.pdf has at least two major merits. First the paper is a brief review of Bayesian and predictive models. Secondly, it critically analyses some virtues of predictive models. We find in agreement with Clark presentation regarding neural economy (energy principle), the presence of different grain sizes in the brain, fundamental aspects of continuity in perception, cognition and action.

What will be next?  First, we need to distinguish between information processing in the biological brain and current artificial models. Second, digital principles borrowed from engineering or machine learning (e.g. prediction error) reflect just tiny parts of multiple computational features expressed by biological neurons. Third, the brain has a specific model of computation by physical (electrical) interaction mediated by molecular changes in neurotransmitters levels (see - neuroelectrodynamics).

Since fragments of information are distributed in various neurons which are densely packed in the brain, then sequential, parallel activation of specific cells is required to integrate information needed for perception or action.  The entire parallelism is bottom up built and the biological brain in all its entirety is the computing machine which exploits  parallel, distributed processing within many  neurons. This specific, continuous (non-Turing) model of computation by electric interaction intrinsically exhibits many features  such as parallelism, fuzziness, fractality  in addition to predictive or Bayesian appearance.  However,

(i)                 The idea of hierarchy in connectionist models has represented an attempt to model anatomical organization. Currently, little experimental data supports the notion of a strict hierarchy in information processing. The presence of various forms of computation at the sub-cellular level and continuous electrical integration of information within neurons do not highlight a strict hierarchy. Related to a specific task or behavior the simultaneous activation of neurons in the brain is required to efficiently integrate information and does not seem to follow a strict anatomical (hierarchical) propagation.

(ii)               From connectivity in proteins or between genes to the association of planets and days of the week, all types of interaction can be approximated by weight type connections. Indeed, the entire idea of weight type connectivity is to present the simplest model. However, this non-specific framework hides various, complex characteristics expressed within different types of interaction in the brain.

(iii)             Even the hypothesis of error prediction Holleman and Schultz, 1998 related to dopaminergic system has to be revised since action potentials are not digital events. Meaningful information is processed and transmitted within every millisecond of spike generation, Aur and Jog, 2010, Aur et al., 2011.

 Therefore, the future of cognitive science seem to stand on specific models which will  have to describe  rich, bioelectrical  interactions (Aur, 2011) that occur within  neurons and  in the  brain.

What will be next? That's a good question and you may find the answer here  http://neuroelectrodynamics.blogspot.com/p/blog-page.html

Aur D.,  Jog, MS, (2010) Neuroelectrodynamics- Understanding The Brain Language , IOS Press 2010, http://dx.doi.org/10.3233/978-1-60750-473-3-i

Aur D., Jog MS, Poznanski, R, 2011, Computing by physical interaction in neurons, Journal of integrative Neuroscience, vol. 10, Issue: 4, , pp. 413-422,

http://www.ncbi.nlm.nih.gov/pubmed/22262533

Aur D., 2011, From Neuroelectrodynamics to Thinking Machines, DOI: 10.1007/s12559-011-9106-3,  Cognitive Computation, 2011, http://www.springerlink.com/content/x1l7388475323758/